Research Of Key Issues In Event Coreference Resolution

As an important component of information extraction,event coreference resolution refers to finding out whether event mentions in the text point to the same event.In the recent years,an issuance of multiple event corpora has provided guarantees for the research of event.However,compared with entity mention,the composition of event mention is more complex and more widely distributed in the documents.Therefore,event coreference resolution is more challenging.At present,most of related work tend to find adaptive features of corpus.The research of within-document event coreference resolution has a certain dependence on specific corpus,and the research of cross-document event resolution is still in its infancy.In response to the status quo and problems above,this paper has focused on three key issue:(1)Targeting at the weal adaptability of the traditional event coreference resolution.A convolution neural network-based event coreference resolution model is constructed by using original context of the event mention,trigger and type of the event mention.In order to verify the validity and generality of the model,the same setting were used to perform experiments in the two public corpora,ACE2005 corpus and KBP2015 corpus.The experimental results show that the proposed model can adapt to the differences between multiple annotations better and it has reached a state-of-art performance in the ACE2005 corpus.(2)Aiming at the problems such as unbalanced instance distribution and the ignorance of intrinsic relations among event mentions in event coreference resolution,a variety of solutions are proposed.The main factors that affect event coreference resolution and the effectiveness of multiple solutions are analysed.(3)Based on the latest KBP2017 corpus,a preliminary exploration was conducted on cross-documents event coreference resolution.Exploting cross-documents coreference resolution into within-document event mention chain detetction and cross-documents event mention chain coreference resolution,and constructing an end-to-end cross-documents event coreference resolution platform.